Deep learning-based bone suppression in chest radiographs using CT-derived features: a feasibility study.

Area of Science:

• Medical Imaging
• Radiology
• Deep Learning

Background:

• Bone suppression in chest X-rays can enhance target localization accuracy in image-guided radiation therapy (IGRT).
• Limited availability of bone-free radiographs restricts the training of bone suppression models.
• This research addresses the scarcity of bone-free datasets by leveraging CT-derived features.

Purpose of the Study:

• To develop a deep learning-based bone suppression method for chest X-rays.
• To reduce the dependency on bone-free datasets for model training.
• To improve accuracy in target localization for IGRT applications.

Main Methods:

• A cascade convolutional neural network (CCNN) was trained using simulated lung digital radiographs (DRs), bone DRs, and bone-free DRs derived from 59 CT scans.
• A three-stage image processing framework involving CT segmentation, DR simulation, and feature expansion was employed.
• External validation on 30 chest radiographs assessed performance using metrics like PSNR, MAE, SSIM, and Spearman's correlation.

Main Results:

• The bone-suppressed radiographs closely matched the reference, achieving high accuracy (MAE=0.0087±0.0030, SSIM=0.8458±0.0317, correlation=0.9554±0.0170, PSNR=20.86±1.60).
• The feature expansion component significantly improved performance across all evaluated metrics.
• Ablation studies confirmed the critical role of feature expansion in the CCNN model's effectiveness.

Conclusions:

• A novel deep learning method using CT-derived features effectively achieves bone suppression in chest X-rays, mitigating the need for extensive bone-free datasets.
• Feature expansion procedures markedly enhanced the model's performance.
• Clinical validation in radiation therapy is recommended to translate this method into practical IGRT applications.